DOCSLIB.ORG

Open Source German Distant Speech Recognition: Corpus and Acoustic Model

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Open Source German Distant Speech Recognition: Corpus and Acoustic Model Open Source German Distant Speech Recognition: Corpus and Acoustic Model Stephan Radeck-Arneth1;2, Benjamin Milde1, Arvid Lange1;2, Evandro Gouvea,ˆ Stefan Radomski1, Max Muhlh¨ auser¨ 1, and Chris Biemann1 1Language Technology Group / 2Telecooperation Group Computer Science Departement Technische Universitat¨ Darmstadt {stephan.radeck-arneth,milde,biem}@cs.tu-darmstadt.de {radomski,max}@tk.informatik.tu-darmstadt.de {egouvea,arvidjl}@gmail.com Abstract. We present a new freely available corpus for German distant speech recognition and report speaker-independent word error rate (WER) results for two open source speech recognizers trained on this corpus. The corpus has been recorded in a controlled environment with three different microphones at a dis- tance of one meter. It comprises 180 different speakers with a total of 36 hours of audio recordings. We show recognition results with the open source toolkit Kaldi (20.5% WER) and PocketSphinx (39.6% WER) and make a complete open source solution for German distant speech recognition possible. Keywords: German speech recognition, open source, speech corpus, distant speech recognition, speaker-independent 1 Introduction In this paper, we present a new open source corpus for distant microphone record- ings of broadcast-like speech with sentence-level transcriptions. We evaluate the corpus with standard word error rate (WER) for different acoustic models, trained with both Kaldi[1] and PocketSphinx[2]. While similar corpora already exist for the German lan- guage (see Table 1), we placed a particular focus on open access by using a permissive CC-BY license and ensured a high quality of (1) the audio recordings, by conducting the recordings in a controlled environment with different types of microphones; and (2) the hand verified accompanying transcriptions. Each utterance in our corpus was simultaneously recorded over different micro- phones. We recorded audio from a sizable number of speakers, targeting speaker inde- pendent acoustic modeling. With a dictionary size of 44.8k words and the best Kaldi model, we are able to achieve a word error rate (WER) of 20.5%. 1.1 Related Work Table 1 shows an overview of current major German speech corpora. Between 1993 and 2000 the Verbmobil [4] project collected around 180 hours of speech for speech transla- tion. The PhonDat [5] corpus was generated during the Verbmobil project and includes 2 Stephan Radeck-Arneth et al. Table 1. Major available corpora containing spoken utterances for the German language. Name Type Size Recorded SmartKom [3] spontaneous, orthographic & 12h synchronized directional & beamformed prosodic Transcription Verbmobil [4] spontaneous, orthographic tran- appr. 180h synchronized close-range & far-field & scription telephone PhonDat [5] spontaneous, orthographic tran- 9.5h close-range scription German Today [6] read and spontaneous, orthographic 1000h single microphone, headset transcription FAU IISAH [7] spontaneous 3.5h synchronized close-range to far-field Voxforge read, orthographic transcription appr. 55h varying microphones, usually headsets Alcohol Language Cor- read, spontaneous, command, or- 300k words close-range, headset pus [8] thographic transcription & control GER-TV1000h [9] read, orthographic transcription appr. 1000h various microphones, broadcast data Our speech corpus read and semi-spontaneous, ortho- appr. 36h x 3 synchronized multiple microphones far- graphic transcription field & beamformed Open Source German Distant Speech Recognition 3 described in [15]. We employed the KisRecord2 software, which supports concurrent recording with multiple microphones. Speakers were presented with text on a screen, one sentence at a time, and were asked to read the text aloud. While the setup is some- what artificial in that reading differs from speaking freely, we avoid the need of tran- scribing the audio and thus follow a more cost-effective approach. For the training part of the corpus, sentences were drawn randomly from three text sources: a set of 175 sentences from the German Wikipedia, a set of 567 utterances from the German section of the European Parliament transcriptions [16] and 177 short com- mands for command-and-control settings. Text sources were chosen because of their free licenses, allowing redistribution of derivatives without restrictions. Test and devel- opment sets were recorded at a later date, with new sentences and new speakers. These have 1028 and 1085 unique sentences from Wikipedia, European Parliament transcrip- tions and crawled German sentences from the Internet, distributed equally per speaker. The crawled German sentences were collected randomly with a focused crawler [17], and were only selected from sentences encountered between quotation marks, which exhibit textual content more typical of direct speech. Unlike in the training set, where multiple speakers read the same sentences, every sentence recorded in the test and de- velopment set is unique, for evaluation purposes. The distance between speaker and microphones was chosen to be one meter, which seems realistic for a business meeting setup where microphones could e.g. be located on a meeting table. There are many more use cases, where a distance of one meter is a sen- sible choice, e.g. in-car speech recognition systems, smart living rooms or plenary ses- sions. The entire corpus is available for the following microphones: Microsoft Kinect, Yamaha PSG-01S, and Samson C01U. The most promising and interesting microphone is the Microsoft Kinect, which is in fact a small microphone array and supports speaker direction detection and beamforming, cf. [15]. We recorded the beamformed and the raw signal simultaneously, however due to driver restrictions both are single channel recordings (i.e. it is not possible to apply our own beamforming on the raw signals). We split the overall data set into training, development and test partitions in such a way that speakers or sentences do not overlap across the different sets. The audio data are available in MS Wave format, one file per sentence per microphone. In addition, for each recorded sentence, an XML file is provided that contains the sentence in original and normalized form (cf. Section 3.2), and the speaker metadata, such as gender, age and region of birth. Table 2 lists the gender and age distribution. Female speakers make up about 30% in all sets and most speakers are aged between 18 and 30 years. For analysis of dialect influence, the sentence metadata also includes the federal state (Bundesland) where the speakers spent the majority of their lives. Despite our corpus being too small for training and testing dialect-specific models, this metadata is collected to support a future training of regional acoustic models. Most speakers are students that grew up and live in Hesse. The statistics related to the number of sentences per speaker for each of the three text sources in the training corpus is given in Table 3. Most sentences were drawn from Wikipedia. On average, utterances from Europarl are longer than sentences from Wikipedia, and speakers encountered more difficulties in reading Europarl utterances 2 http://kisrecord.sourceforge.net 4 Stephan Radeck-Arneth et al. Table 3. Mean and standard deviation of training set Table 2. Speaker gender and age dis- sentences read per person per text source and total tribution audio recording times for each microphone Gender Train Dev Test male 105 12 13 Corpus Mean ± StD female 42 4 4 Wikipedia 35 ± 12 Europarl 8 ± 3 Age Train Dev Test Command 18 ± 23 41–50 2 0 1 31–40 17 17 2 Microphone Dur. Train/Dev/Test (h) 21–30 108 13 10 Microsoft Kinect 31 / 2 / 2 18–20 20 1 4 Yamaha PSG-01S 33 / 2 / 2 Samson C01U 29 / 2 / 2 because of their length and domain specificity. Recordings were done in sessions of 20 minutes. Most speakers participated in only one session, some speakers (in the training set) took part in two sessions. Table 3 also compares the audio recording lengths of the three main microphones for each dataset. Occasional outages of audio streams oc- curred during the recording procedure, causing deviations in recording length for each microphone. By ensuring that the training and the dev/test portions have disjoint sets of speakers and textual material, this corpus is perfectly suited for examining approaches to speaker-independent distant speech recognition for German. 3 Experiments In this section, we show how our corpus can be used to generate a speaker-independent model in PocketSphinx and Kaldi. We then compare both speech recognition toolkits using the same language model and pronunciation dictionary in terms of word error rate (WER) on our heldout data from unseen speakers (development and test utter- ances). Our Kaldi recipe has been released on Github3 as a package of scripts, which support fully automatic generation of all acoustic models with automatic downloading and preparation of all needed resources, including phoneme dictionary and language model. This makes the Kaldi results easily reproducible. 3.1 Phoneme dictionary As our goal is to train a German speech recognizer using only freely available sources, we have not relied on e.g. PHONOLEX4, which is a very large German pronunci- ation dictionary with strict licensing. We compiled our own German pronunciation dictionary using all publicly available phoneme lexicons at the Bavarian Archive for speech signals (BAS)5 and using the MaryTTS [18] LGPL-licensed pronunciation dic- tionary, which has 26k entries. Some of the publicly available BAS dictionaries, like the one for the Verbmobil [4] corpus, also contain pronunciation variants which were 3 https://github.com/tudarmstadt-lt/kaldi-tuda-de 4 https://www.phonetik.uni-muenchen.de/Bas/BasPHONOLEXeng.html 5 ftp://ftp.bas.uni-muenchen.de/pub/BAS/ Open Source German Distant Speech Recognition 5 included. The final dictionary covers 44.8k unique German words with 70k total en- tries, with alternate pronunciations for some of the more common words. Stress mark- ers in the phoneme set were grouped with their unstressed equivalents in Kaldi using the extra_questions.txt file and were entirely removed for training with CMU Sphinx.
Load more

Recommended publications
  • Speech Recognition in Java Breandan Considine Jetbrains, Inc
    Speech Recognition in Java Breandan Considine JetBrains, Inc. Automatic speech recognition in 2011 Automatic speech recognition in 2015 What happened? • Bigger data • Faster hardware • Smarter algorithms Traditional ASR • Requires lots of handmade feature engineering • Poor results: >25% WER for HMM architectures State of the art ASR • <10% average word error on large datasets • DNNs: DBNs, CNNs, RBMs, LSTM • Thousands of hours of transcribed speech • Rapidly evolving field • Takes time (days) and energy (kWh) to train • Difficult to customize without prior experience Free / open source • Deep learning libraries • C/C++: Caffe, Kaldi • Python: Theano, Caffe • Lua: Torch • Java: dl4j, H2O • Open source datasets • LibriSpeech – 1000 hours of LibriVox audiobooks • Experience is required Let’s think… • What if speech recognition were perfect? • Models are still black boxes • ASR is just a fancy input method • How can ASR improve user productivity? • What are the user’s expectations? • Behavior is predictable/deterministic • Control interface is simple/obvious • Recognition is fast and accurate Why offline? • Latency – many applications need fast local recognition • Mobility – users do not always have an internet connection • Privacy – data is recorded and analyzed completely offline • Flexibility – configurable API, language, vocabulary, grammar Introduction • What techniques do modern ASR systems use? • How do I build a speech recognition application? • Is speech recognition accessible for developers? • What libraries and frameworks exist
    [Show full text]
  • Multimodal HALEF: an Open-Source Modular Web-Based Multimodal Dialog Framework
    Multimodal HALEF: An Open-Source Modular Web-Based Multimodal Dialog Framework Zhou Yu‡†, Vikram Ramanarayanan†, Robert Mundkowsky†, Patrick Lange†, Alexei Ivanov†, Alan W Black‡ and David Suendermann-Oeft† Abstract We present an open-source web-based multimodal dialog framework, “Multimodal HALEF”, that integrates video conferencing and telephony abilities into the existing HALEF cloud-based dialog framework via the FreeSWITCH video telephony server. Due to its distributed and cloud-based architecture, Multimodal HALEF allows researchers to collect video and speech data from participants in- teracting with the dialog system outside of traditional lab settings, therefore largely reducing cost and labor incurred during the traditional audio-visual data collection process. The framework is equipped with a set of tools including a web-based user survey template, a speech transcription, annotation and rating portal, a web visual processing server that performs head tracking, and a database that logs full-call audio and video recordings as well as other call-specific information. We present observations from an initial data collection based on an job interview application. Finally we report on some future plans for development of the framework. 1 Introduction and Related Work Previously, many end-to-end spoken dialog systems (SDSs) used close-talk micro- phones or handheld telephones to gather speech input [3] [22] in order to improve automatic speech recognition (ASR) performance of the system. However, this lim- its the accessibility of the system. Recently, the performance of ASR systems has improved drastically even in noisy conditions [5]. In turn, spoken dialog systems are now becoming increasingly deployable in open spaces [2].
    [Show full text]
  • Systematic Review on Speech Recognition Tools and Techniques Needed for Speech Application Development
    INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 03, MARCH 2020 ISSN 2277-8616 Systematic Review On Speech Recognition Tools And Techniques Needed For Speech Application Development Lydia K. Ajayi, Ambrose A. Azeta, Isaac. A. Odun-Ayo, Felix.C. Chidozie, Aeeigbe. E. Azeta Abstract: Speech has been widely known as the primary mode of communication among individuals and computers. The existence of technology brought about human computer interface to allow human computer interaction. Existence of speech recognition research has been ongoing for the past 60 years which has been embraced as an alternative access method for individuals with disabilities, learning shortcomings, Navigation problems etc. and at the same time allow computer to identify human languages by converting these languages into strings of words or commands. The ongoing problem in speech recognition especially for under-resourced languages has been identified to be background noise, speed, recognition accuracy, etc. In this paper, in order to achieve a higher level of speed, recognition accuracy (word or sentence accuracy), manage background noise, etc. we have to be able to identify and apply the right speech recognition algorithms/techniques/parameters/tools needed ultimately to develop any speech recognition system. The primary objective of this paper is to make summarization and comparison of some of the well-known methods/algorithms/techniques/parameters identifying their steps, strengths, weaknesses, and application areas in various stages of speech recognition systems as well as for the effectiveness of future services and applications. Index Terms: Speech recognition, Speech recognition tools, Speech recognition techniques, Isolated speech recognition, Continuous speech recognition, Under-resourced languages, High resourced languages.
    [Show full text]
  • Hidden Voice Commands
    Hidden Voice Commands Nicholas Carlini∗ Pratyush Mishra University of California, Berkeley University of California, Berkeley Tavish Vaidya Yuankai Zhang Micah Sherr Georgetown University Georgetown University Georgetown University Clay Shields David Wagner Wenchao Zhou Georgetown University University of California, Berkeley Georgetown University Abstract command may recognize it as an unwanted command Voiceinterfaces are becoming more ubiquitous and are and cancel it, or otherwise take action. This motivates now the primary input method for many devices. We ex- the question we study in this paper: can an attacker cre- plore in this paper how they can be attacked with hidden ate hidden voice commands, i.e., commands that will be voice commands that are unintelligible to human listen- executed by the device but which won’t be understood ers but which are interpreted as commands by devices. (or perhaps even noticed) by the human user? We evaluate these attacks under two different threat The severity of a hidden voice command depends upon models. In the black-box model, an attacker uses the what commands the targeted device will accept. De- speech recognition system as an opaque oracle. We show pending upon the device, attacks could lead to informa- that the adversary can produce difficult to understand tion leakage (e.g., posting the user’s location on Twitter), commands that are effective against existing systems in cause denial of service (e.g., activating airplane mode), the black-box model. Under the white-box model, the or serve as a stepping stone for further attacks (e.g., attacker has full knowledge of the internals of the speech opening a web page hosting drive-by malware).
    [Show full text]
  • German Speech Recognition: a Solution for the Analysis and Processing of Lecture Recordings
    German Speech Recognition: A Solution for the Analysis and Processing of Lecture Recordings Haojin Yang, Christoph Oehlke, Christoph Meinel Hasso Plattner Institut (HPI), University of Potsdam P.O. Box 900460, D-14440 Potsdam e-mail: Haojin.Yang, Meinel @hpi.uni-potsdam.de; [email protected] { } Abstract—Since recording technology has become more ro- Therefore their training dictionary can not be extended or bust and easier to use, more and more universities are taking optimized periodically. [8] proposed a solution for creating the opportunity to record their lectures and put them on the an English speech corpus basing on lecture audio data. But, Web in order to make them accessable by students. The auto- matic speech recognition (ASR) techniques provide a valueable they have not dealt with a complete speech recognition source for indexing and retrieval of lecture video materials. In system. [7] introduced a spoken document retrieval system this paper, we evaluate the state-of-the-art speech recognition for Korean lecture search. Their automatically generated software to find a solution for the automatic transcription of search index table is based on lecture transcriptions. How- German lecture videos. Our experimental results show that ever, they do not consider the recognition of topic-related the word error rates (WERs) was reduced by 12.8% when the speech training corpus of a lecturer is increased by 1.6 hours. technical foreign words which are important for keyword- search. Overall, most of these lecture recognition systems Keywords-automatic speech recognition; e-learning; multi- have a low recognition accuracy, the WERs of audio lec- media retrieval; recorded lecture videos.
    [Show full text]
  • Bootstrapping Development of a Cloud-Based Spoken Dialog System in the Educational Domain from Scratch Using Crowdsourced Data
    Research Report ETS RR–16-16 Bootstrapping Development of a Cloud-Based Spoken Dialog System in the Educational Domain From Scratch Using Crowdsourced Data Vikram Ramanarayanan David Suendermann-Oeft Patrick Lange Alexei V. Ivanov Keelan Evanini Zhou Yu Eugene Tsuprun Yao Qian May 2016 ETS Research Report Series EIGNOR EXECUTIVE EDITOR James Carlson Principal Psychometrician ASSOCIATE EDITORS Beata Beigman Klebanov Donald Powers Senior Research Scientist - NLP ManagingPrincipalResearchScientist Heather Buzick Gautam Puhan Research Scientist Principal Psychometrician Brent Bridgeman John Sabatini Distinguished Presidential Appointee ManagingPrincipalResearchScientist Keelan Evanini Matthias von Davier Managing Senior Research Scientist - NLP Senior Research Director Marna Golub-Smith Rebecca Zwick Principal Psychometrician Distinguished Presidential Appointee Shelby Haberman Distinguished Presidential Appointee PRODUCTION EDITORS Kim Fryer Ayleen Stellhorn Manager, Editing Services Senior Editor Since its 1947 founding, ETS has conducted and disseminated scientific research to support its products and services, and to advance the measurement and education fields. In keeping with these goals, ETS is committed to making its research freely available to the professional community and to the general public. Published accounts of ETS research, including papers in the ETS Research Report series, undergo a formal peer-review process by ETS staff to ensure that they meet established scientific and professional standards. All such ETS-conducted peer reviews are in addition to any reviews that outside organizations may provide as part of their own publication processes. Peer review notwithstanding, the positions expressed in the ETS Research Report series and other published accounts of ETS research are those of the authors and not necessarily those of the Officers and Trustees of Educational Testing Service.
    [Show full text]
  • Using Open Source Speech Recognition Software Without an American Accent
    Using open source speech recognition software without an American accent. The field Julius: C, strange BSDish license, Japanese HTK: (HMM (Hidden Markov Model) Toolkit) C, non-commercial license Kaldi: C++ library, Apache 2.0 (F.U. HTK) CMU Sphinx Family: C, Java, BSD-ish Smaller possibly dormant projects Shout: C++, GPL RWTH ASR: C++, non-commercial, German Iatros: C, GPL3, Spanish Carnegie Mellon University Sphinx Sphinx C 1980s Sphinx 2 C 1990s -------. | Sphinx 3 C 2000s | | Sphinx 4 Java | | PocketSphinx C Sphinx 3, 4, and PocketSphinx share tools. Sphinx development model PhD driven: 1. Have a moderately good idea 2. Write code that “proves” your idea 3. Submit thesis 4. $$$* * $$$ refers to a job at Nuance, Apple, Microsoft, Google, IBM,... The “Coral Reef” development model. What Americans do: sudo aptitude install \ pocketsphinx-hmm-wsj1 \ pocketsphinx-lm-wsj \ python-pocketsphinx \ # or gstreamer0.10-pocketsphinx from pocketsphinx import Decoder HMM = "/usr/share/pocketsphinx/model/hmm/wsj1/" LM = "/usr/share/pocketsphinx/model/"\ "lm/wsj/wlist5o.3e-7.vp.tg.lm.DMP" DICT = "/usr/share/pocketsphinx/model/"\ "lm/wsj/wlist5o.dic" decoder = Decoder(hmm=HMM, lm=LM, dict=DICT) fh = open("speech.wav") fh.seek(44) # skip the WAV header decoder.decode_raw(fh) print decoder.get_hyp() # short for hypothesis A good reference is David Huggins-Daines’ live-coding PyCon2010 lightning talk. Everything goes wrong and he still finishes in under 4 minutes. The configuration lines: HMM = ".../hmm/wsj1/" LM = ".../lm/wlist5o.3e-7.vp.tg.lm.DMP" DICT = ".../lm/wsj/wlist5o.dic" HMM Acoustic model (Hidden Markov Model) LM Language model DICT Pronunciation dictionary The Acoustic model matches sounds to phoneme probabilities.
    [Show full text]
  • Automatic Derivation of Grammar Rules Used for Speech Recognition in Dialogue Systems
    Masaryk University Faculty of Informatics Automatic Derivation of Grammar Rules Used for Speech Recognition in Dialogue Systems Master’s Thesis Bc. Klára Kufová Brno, Spring 2018 This is where a copy of the official signed thesis assignment and a copy of the Statement of an Author is located in the printed version of the document. Declaration Hereby I declare that this paper is my original authorial work, which I have worked out on my own. All sources, references, and literature used or ex- cerpted during elaboration of this work are properly cited and listed in com- plete reference to the due source. Bc. Klára Kufová Advisor: Mgr. Luděk Bártek, Ph.D. i Acknowledgements First and foremost, I would like to acknowledge the thesis advisor, Mgr. Luděk Bártek, Ph.D., of the Faculty of Informatics at Masaryk University, for his valuable ideas, constructive advice, and the time dedicated to our consultations. My sincere thanks also go to both my current and former colleagues at Red Hat, especially to Mgr. Milan Navrátil and Ing. Radovan Synek, who helped me tremendously during both the implementation and writing of this thesis. Last but not least, I would like to express my sincere gratitude to my parents, grandparents, my brother, and to my fiancé, who have supported and encouraged me throughout the course of my studies. iii Abstract The thesis deals with the process of building an adaptive dialogue system, which is capable of learning new grammar rules used for automatic speech recognition based on past conversations with real users. The subsequent abil- ity to automatically reduce the unused rules within a grammar is proposed and implemented as well.
    [Show full text]
  • End User License Agreement
    End User License Agreement End User License Agreement 14 January 2016 Contact: LumenVox, LLC Edward Miller, CEO 3615 Kearny Villa Road, Suite 202, San Diego, CA 92123 USA Telephone: +1 (858) 707-7700 Email: [email protected] Website: www.LumenVox.com Version 15.0 Page 1 January 2016 End User License Agreement This License Agreement between LumenVox LLC (Hereinafter “LumenVox”) and You (Hereinafter “You”) constitutes the terms and conditions for Your use of the Software Product in consideration of the license fees that You have paid. The Software Product(s) provided with this agreement are licensed, not sold, to You for use only under the terms of this License Agreement, unless a Software Product is accompanied by a separate license agreement, in which case the terms of that separate license agreement will govern, subject to Your prior acceptance of that separate license agreement. Please read this Agreement carefully. By downloading, installing, copying, or otherwise using the Software, You agree to be bound by the terms and conditions of this Agreement and become a party to this Agreement. If You do not agree with all of the terms and conditions of this Agreement, do not download, install, copy or otherwise use the Software. You may, however, request a return for a full refund within 30 days of your purchase. THE SOFTWARE PRODUCT IS PROTECTED UNDER U.S. AND INTERNATIONAL COPYRIGHT LAWS, AS WELL AS OTHER INTELLECTUAL PROPERTY LAWS AND TREATIES. 1. DEFINITIONS. The following definitions apply to this Licensing Agreement: 1.1. “Application Context” means the particular fields of use and the purpose and manner of using the Software permitted under this Agreement, as specified below.
    [Show full text]
  • Scientific Learning Reading Assistant™: Cmu Sphinx Technology in a Commercial Educational Software Application
    SCIENTIFIC LEARNING READING ASSISTANT™: CMU SPHINX TECHNOLOGY IN A COMMERCIAL EDUCATIONAL SOFTWARE APPLICATION Valerie L. Beattie, Ph.D. Scientific Learning Corporation 300 Frank H. Ogawa Plaza, Suite 600, Oakland, CA 94612 www.scilearn.com [email protected] ABSTRACT research[6]. However, given public school class sizes and budget constraints, it is not possible for teachers or other The Reading Assistant™ software application is a reading trained professionals to provide sufficient practice with this tutor designed to provide the support and practice needed to kind of one-on-one help and feedback to students while improve reading skills. CMU Sphinx speech recognition they read. Speech recognition technology has the potential technology is used to provide the guided oral reading to fill this gap, but the challenges are also considerable. On experience that is key to building reading fluency and the one hand, the speech recognition software must ability. The application listens and follows along as a accommodate the acoustic variability from a wide range of student reads the displayed text, providing feedback and user ages, accents, reading rates, and noise conditions. At help if the student gets stuck or makes an error. the same time the application must strive to identify In many ways this application of speech recognition ‘undesirable’ variations in the user’s reading of the text, i.e. technology is different from typical commercial reading errors. applications, and we have made many modifications to the The development of the Reading Assistant application CMU Sphinx technology to adapt it to the needs of the began in 2000 with the foundation of Soliloquy Learning, Reading Assistant application.
    [Show full text]
  • Making an Effective Use of Speech Data for Acoustic Modeling
    Making an Effective Use of Speech Data for Acoustic Modeling Rong Zhang CMU-LTI-07-016 Language Technologies Institute School of Computer Science Carnegie Mellon University 5000 Forbes Ave., Pittsburgh, PA 15213 www.lti.cs.cmu.edu Thesis Committee: Alexander I. Rudnicky, Chair Richard Stern Tanja Schultz Karthik Visweswariah Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy In Language and Information Technologies © 2007, Rong Zhang Abstract Automatic recognition of continuous speech has been acknowledged as one of the most challenging problems today. The performance of a continuous speech recognition system highly depends on the availability of sufficient speech data and transcripts of good quality. In most cases, however, carefully prepared in-domain data is not easy to obtain because collecting a large amount of transcribed speech data is normally a time-consuming and expensive process. The acoustic model trained without the support of sufficient training data is less capable in handling the complexity and variability of human speech, and thus performs poorly in real world application. This raises us the questions such as how to effectively exploit the given training data to improve the performance of recognition systems, and how to explore error-prone but informative data sources and incorporate them into acoustic model training. This thesis summarizes our efforts in investigating solutions to address the above issues. The work can be divided into two parts. We first investigate Boosting algorithm, an ensemble based supervised training approach, which iteratively creates multiple acoustic models with complementary error patterns by manipulating the distribution of training data.
    [Show full text]
  • Title, Modify
    Muhammad Salman Ishaq Voice Activity Detection and Garbage Modelling for a Mobile Automatic Speech Recognition Application School of Electrical Engineering Thesis submitted for examination for the degree of Master of Science in Technology. Espoo 23.1.2017 Thesis supervisor: Prof. Mikko Kurimo Thesis advisors: Matti Varjokallio Leo Hämäläinen aalto university abstract of the school of electrical engineering master’s thesis Author: Muhammad Salman Ishaq Title: Voice Activity Detection and Garbage Modelling for a Mobile Automatic Speech Recognition Application Date: 23.1.2017 Language: English Number of pages: 8+55 Department of Signal Processing and Acoustic Professorship: Speech Recognition Code: S3013 Supervisor: Prof. Mikko Kurimo Advisors: Matti Varjokallio, Leo Hämäläinen Recently, state-of-the-art automatic speech recognition systems are used in various industries all over the world. Most of them are using a customized version of speech recognition system. The need for different versions arise due to different speech commands, lexicon, language and distinct work environment. It is essential for a speech recognizer to provide accurate and precise outputs in every working environment. However, the performance of a speech recognizer degrades quickly when noise intermingles with a work environment and also when out-of-vocabulary (OOV) words are spoken to the speech recognizer. This thesis consists of three different tasks which improve an automatic speech recognition application for mobile devices. The three tasks include building of a new acoustic model, improving the current voice activity detection and garbage modelling of OOV words. In this thesis, firstly, a Finnish acoustic model is trained for a company called Devoca Oy. The training data was recorded from different warehouse environments to improve the real-world speech recognition accuracy.
    [Show full text]